DRDaniel Roberts
MayaWorld · AI-orchestrated delivery

MayaWorld, a production game shipped by a team of AI agents I directed

I was the lead orchestrator. Four frontier models, each routed to what it does best, built a real-time multiplayer 3D world end-to-end: a custom browser engine, live multiplayer, an avatar economy, and a 56-quest campaign. Two weeks of nights, concept to production. The point isn't the game; it's the operating model, this is how I take anything 0→1 and get it to market in weeks, solo.

2 wksnights-only, 0→live
4models orchestrated
~20klines shipped
31playable scenes
Orchestration map

Four models, one orchestrator.

OROrchestrator, me
Owned the roadmap and the quality gate. Decompose features into model-sized tasks · route each task to the model best suited to it · integrate the pieces · verify every change in the running product before it counts · ship nightly. The models do the work; the orchestration is the job.
M1Fable, the manager
Kept the agent team working. Dispatched tasks, tracked status, unblocked stalled work, and kept the build moving between my check-ins, the layer that turned four models into one team.
M2Claude Opus 4.8, coding agent
Wrote the system. The custom 3D engine (React + Three.js/WebGL), Socket.IO real-time multiplayer, gameplay systems, and long multi-file features, verified in the running product.
M3ChatGPT 5.6, documentation
The paper trail. Specs, design docs, and the white paper, the shared record that kept a four-model team coherent across two weeks of nightly builds.
M4Kimi K3, pipeline & refactor
Scale work. Asset-pipeline scripting (200+ CC0 models converted to glTF) and large, cost-efficient refactor passes over long context.

Same discipline as my platform work: decomposed scope, evidence before merge, shipped increments. Different workforce.

The product

Built against a stakeholder who tells me the truth.

My daughter wanted a game with her ideas in it, so she ran product. She wrote the feature list, played every build, and rejected most of them: the fastest, most honest feedback loop I've ever shipped against. Requirements changed weekly, scope was cut ruthlessly, and the customer was ten and unimpressed by technical excuses. What shipped: a shared roleplay town, five co-op monster-chase games, a pet-battle RPG with hybrid breeding and evolutions, a 56-quest campaign across themed 3D biomes, plus racing, dungeon, survival and arena modes, running on a custom engine in any browser, no install.

The demo above is a sandboxed public build with sample characters, no family accounts or personal data are reachable from it. Live, in production, with one extremely engaged user.